EastVoice · AI · 2026-09-24 · Issue #77

Editor's Note: Xiaomi's MiMo-V3 lands with a new architecture and the HySparse 2 component behind it; iFlytek ships Spark-ASR-2.0 into its keyboard; Alibaba turns its Qwen platform into a full MaaS and agent stack. Alibaba rotated the executive running its flagship model on the eve of Yunqi. Musk told CCTV that Chinese models already deliver near world-leading performance per unit of compute, and that the chip gap closes in two to three years. Peng Zhihui starts selling a ¥20,000 / ~$2,800 robot you can program at home.

Commentary: OpenAI and Anthropic cut prices 90 minutes apart, and Chinese labs answered with architecture, hardware and internal tooling. Cheap tokens are table stakes now. The fight is over who owns the device and the data that feeds it.

---

I. AI Foundation Models

1. Xiaomi's MiMo-V3 Arrives With a New Architecture and HySparse 2

• Summary: ITHome reported on Sept 23 that Xiaomi's Luo Fuli (罗福莉) confirmed MiMo-V3 runs on a new architecture, with the core HySparse 2 component released the same day. Xiaomi's bet is that sparsity, not parameter count, is where the next efficiency gain comes from.
• Source: ITHome · 2026-09-23
• Editor's Take: Everyone can buy the same GPUs, so the differentiator has to sit in the architecture. If HySparse 2 survives production traffic, Xiaomi owns something no cluster can buy.

2. iFlytek Releases Spark-ASR-2.0 Speech Recognition Model

• Summary: iFlytek (科大讯飞) launched Spark-ASR-2.0 on Sept 23, combining non-autoregressive decoding with an LLM-enhanced autoregressive path, plus dynamic context injection. It goes live in the iFlytek keyboard on Sept 24, with the largest claimed gains on mixed Chinese-English speech.
• Source: ITHome · 2026-09-23
• Editor's Take: Ambient dictation is where Chinese speech models actually collect revenue. For anyone typing WeChat messages, mixed-language accuracy matters more than a word error rate on a benchmark.

3. Alibaba Rebuilds the Qwen Platform Around MaaS and Agents

• Summary: At the Yunqi conference on Sept 22, Alibaba detailed an upgrade spanning model services, agent services and industry AI solutions, presented by the group's strategy VP who runs the ATH business group's MaaS line. The pitch covers everything from hosted model access to vertical deployments.
• Source: QbitAI · 2026-09-23
• Editor's Take: Alibaba's answer to a token price war is to sell the whole stack instead of the cheapest call. Platform lock-in is the margin that survives when API prices halve.

4. Ant Group Launches Mishuan No.1, a Confidential-Computing Platform for Enterprises

• Summary: Ant Group's confidential-computing unit unveiled Mishuan No.1 (密算一号) at Yunqi, an enterprise-grade trusted intelligent cloud platform that connects cloud models, agents and compute over confidential computing, with an invitation-only first cohort. The claim is that corporate data stays usable but invisible, controllable and metered across institutions.
• Source: Leiphone · 2026-09-23
• Editor's Take: Cross-institution data deals die on liability, not on cryptanalysis. Ant is selling the audit trail as much as the encryption.

5. Zhipu AI Confronts the Hardest Trust Test in Its History

• Summary: Huxiu published an analysis on Sept 23 tracing Zhipu AI's crisis to a 313MB encrypted package that failed to upload 564 times, sitting next to a plaintext manifest of more than 40,000 files, over 80% of them from .git directories. The fallout landed days after Anthropic named Zhipu in a 154-page safety report on Chinese labs.
• Source: Huxiu · 2026-09-23
• Editor's Take: The leak is a security failure, and a security failure found this close to a US report naming you is a sales problem. Enterprise buyers will want evidence, not statements.

6. TypeSafe AI's Jev Reframes 'Decision AI'

• Summary: QbitAI profiled Jev on Sept 23. The team behind it, with OpenAI alumni on staff, skipped generation and reasoning to work on simulation and decision-making for complex real-world problems, and the article compares it with Decitron to separate two things that get filed under the same label.
• Source: QbitAI · 2026-09-23
• Editor's Take: Decision AI usually means a narrow simulator wrapped in a good demo. The number worth checking is whether it beats a tuned classical planner on the same task at the same cost.

7. Musk Tells CCTV China's Models Are Near the Top on Performance Per Unit of Compute

• Summary: In a CCTV Finance interview on Sept 23, Musk said China's large models are excellent overall and deliver nearly world-leading performance per unit of compute. He expects the compute gap to close in roughly two to three years as lithography and chip manufacturing improve.
• Source: ITHome · 2026-09-23
• Editor's Take: From the largest buyer of Nvidia silicon, efficiency praise is a statement about export controls. The interesting part is what happens to the market when the gap actually closes.

8. Alibaba Moves Its Flagship Model Chief Off the Yunqi Stage

• Summary: TMTPost noted on Sept 23 that Zhou Jingren (周靖人), a fixture on the Yunqi main stage for three years, was absent from this year's speaker and VIP lists, replaced by the head of the Tongyi Qianwen project appointed the day before the conference opened.
• Source: TMTPost · 2026-09-23
• Editor's Take: Changing who fronts the flagship model is what companies do when priorities move faster than launches. This year's Yunqi announcements will show whether it was a personnel fix or a strategy shift.

9. A Bengio Protege Leaves DeepMind for OpenAI

• Summary: A researcher who worked on Gemini 2.5 and 3, the Genie world models and the Sima embodied agent at DeepMind moved to OpenAI on Sept 16, per a Huxiu account published Sept 23.
• Source: Huxiu · 2026-09-23
• Editor's Take: Talent flows toward whoever owns the next capability curve, and world models plus embodied agents sit on that curve. DeepMind can absorb the loss. The direction of the flow is the story.

10. Chinese Tech Press Reads the Global Price War as a Land Grab

• Summary: Chinese outlets covered Sept 23 as one event. Anthropic shipped Claude Opus 5.5 at 40% lower running cost, and OpenAI answered 90 minutes later with GPT-6 Sol and GPT-6 Luna at half the previous API price. Huxiu framed both moves as defending enterprise accounts rather than pushing the capability frontier.
• Source: Huxiu · 2026-09-23
• Editor's Take: Two labs cutting prices within 90 minutes says what the token business looks like now. Capacity is abundant, differentiation is thin, and the money goes into distribution.

---

II. AI Software & Applications

1. Doubao Phones and QwenBook Tablets Turn Assistants Into a Hardware Fight

• Summary: After ByteDance (字节跳动) shipped a second-generation Doubao phone, Alibaba showed QwenBook, a tablet still in development from its Wuying team, at Yunqi on Sept 22, per LatePost and a Huxiu summary. Both companies are moving from app-layer assistants toward owning the device.
• Source: Huxiu · 2026-09-23
• Editor's Take: An assistant that lives inside someone else's app can be evicted by a system update. Building the device is the expensive way to buy back the default position.

2. Leapmotor Says 40-Plus Internal AI Systems Cut Build Time by 90%

• Summary: Leiphone reported Sept 23 that Leapmotor (零跑汽车) built more than 40 internal AI systems on Tencent's WorkBuddy, covering software quality, engineering digitalization and lab management. One stamping-die system spanning 12 modules went live in 10 person-days, against roughly 100 person-days under the traditional IT process.
• Source: Leiphone · 2026-09-23
• Editor's Take: The 10x number came from tools built by business staff outside central IT. That is a governance problem wearing a productivity headline, and the shadow systems will surface at the first audit.

3. AI Rewrites the Hiring Bar in Chinese Game Studios

• Summary: Huxiu reported on Sept 23 that AI has reset entry-level expectations in games and short drama production. One studio now asks art graduates for 5 to 10 review-ready concept pieces a day, then rough 3D models within two days.
• Source: Huxiu · 2026-09-23
• Editor's Take: The output bar moved faster than the training pipeline that used to set it. Junior artists now compete against their own tools on volume, which is a strange place to start a career.

4. A Meta Agent Rally in the US Lifted Tencent, and Chinese Media Asked Why

• Summary: TMTPost examined on Sept 23 why Tencent shares rose while Meta's Muse dominated headlines. Meta's personal agent drew 730,000 downloads in five days and 2.5 million in thirteen, topped the US App Store free chart, and reported 642,000 US daily active users.
• Source: TMTPost · 2026-09-23
• Editor's Take: Chinese investors read a US consumer agent hitting number one as a signal about WeChat's own agent timetable. That is how the market prices a product Tencent has not shipped.

5. Qualcomm Says Apps Are Giving Way to Agents, With the Phone Still at the Center

• Summary: Qualcomm's senior VP for mobile business, Chris Patrick, said in comments covered Sept 23 that smartphones remain the core of personal computing, with usage shifting from standalone apps toward AI agents that act on the user's behalf.
• Source: ITHome · 2026-09-23
• Editor's Take: A chip vendor announcing the end of the app economy is telling you where the next socket is. Agent traffic is harder to attribute to a single app and easier to bill per device.

---

III. Humanoid Robots

1. Peng Zhihui Sells a ¥20,000 Robot That Can Be Humanoid or Quadruped

• Summary: QbitAI reported Sept 23 that Shangwei New Material (上纬新材), chaired by Peng Zhihui (稚晖君), is now selling a robot at ¥20,000 / ~$2,800 that can be configured as a humanoid or a quadruped and is open to developers.
• Source: QbitAI · 2026-09-23
• Editor's Take: Pricing a developer robot below a good workstation is how you get thousands of builders writing against your stack. Volume on that tier decides who collects the embodied data.

2. Qiyuan Robots Puts the Q1 and T1 Consumer Humanoids on Sale

• Summary: TMTPost reported Sept 23 that Qiyuan (启元机器人) began selling the Q1, described as the first small humanoid with full-body force control, and the transformable personal robot T1 on Sept 20, promising shipping at launch and offering customization of appearance and voice.
• Source: TMTPost · 2026-09-23
• Editor's Take: Celebrity endorsements buy attention from families, not retention. The number worth watching is how many units are still switched on after the novelty month.

3. Benmo Dynamics Takes Its Robot-Joint Business to a Hong Kong IPO

• Summary: TMTPost reported Sept 23 that Benmo Dynamics (本末动力) opened its Hong Kong main board offering on Sept 21, with trading expected Sept 29. Direct-drive joint modules are the revenue base, and robot vacuum makers are the main customers.
• Source: TMTPost · 2026-09-23
• Editor's Take: A joint supplier whose revenue comes from vacuum cleaners is being valued on embodied AI. The prospectus will show how much of the story is order book.

4. Musk Predicts a Billion Humanoid Robots Within a Decade

• Summary: In the same CCTV interview on Sept 23, Musk said everyone will eventually have a personal robot that can care for aging parents, watch children and work as a one-on-one tutor, and predicted at least 1 billion humanoid robots within ten years, producing a wave of one-person companies.
• Source: ITHome · 2026-09-23
• Editor's Take: A billion units in ten years implies actuator and battery supply chains that do not exist yet. Treat it as a demand forecast, and nothing more.

---

IV. Autonomous Driving

1. Tesla Brings Grok Bot to North American Cars

• Summary: ITHome reported Sept 23 that Tesla's North American fleet now runs Grok Bot on the head unit, letting drivers order coffee, send email and handle errands hands-free by voice, beyond the in-car questions and climate control it handled before.
• Source: ITHome · 2026-09-23
• Editor's Take: Turning the cabin into an agent surface is cheap to ship and adds real value. A misheard verb now has spending power behind it, and that warranty question is unanswered.

2. Deepal's S07 Puts ByteDance's Doubao Model in the Cabin

• Summary: Deepal (深蓝汽车) announced a partnership with Volcano Engine on Sept 23 to build cockpit features on the Doubao model, starting with the new S07 AI Laser Edition: natural-language route planning, voice control from outside the vehicle, and fuzzy requests instead of fixed commands.
• Source: ITHome · 2026-09-23
• Editor's Take: Accepting vague phrasing is the part users actually notice, and Chinese automakers are shipping it faster than anyone else. The cockpit is the country's busiest agent deployment.

---

V. World Models / Physical AI

1. PixVerse R2 Pushes Real-Time World Models Toward General Use

• Summary: QbitAI profiled PixVerse R2 on Sept 23, arguing that real-time generative world models are leaving the demo stage, with finer control and higher fidelity inside the same real-time loop.
• Source: QbitAI · 2026-09-23
• Editor's Take: Real-time generation and fine control rarely arrive together. Whether the pair holds under interactive load is the question a demo cannot answer.

2. The First Team to Automate AI Research May Not Own the Best Model

• Summary: TMTPost published an analysis on Sept 23 arguing that progress in self-improving systems depends on research infrastructure and evaluation loops as much as on raw model strength. The framing: in 2026, AI has moved from an object of research to a participant in it.
• Source: TMTPost · 2026-09-23
• Editor's Take: Evaluation loops are the bottleneck and the least glamorous place to invest. Whichever lab automates its own eval pipeline first gets compounding returns on everything after.

---

物界前沿 | EastVoice